Using slighly different solver options should fix most of the errors when the test problems are solved with COSMO.jl.
Current errors:
LPs
The errors disappear when the number of maximum iterations is increased to ~10000. As COSMO is using only first-order information and the tolerances are set to 1e-6 it will need more iterations than the default max_iter = 2500
sdp_Partial_trace: Again caused by COSMO trying to decompose the LMI constraint. This is actually an interesting edge case that I haven't considered. The problem has a cascaded LMI constraint in its conic form: A = [B 0; 0 0] >= 0 where B >= 0. I am not sure atm how to efficiently check for these cases. That's why I turned the decomposition off for this testset, i.e. decomp = false.
sdp_sdp_constraints: The problem suffers from very slow convergence. That's why I slightly increased the tolerances to 5e-7 and max_iter = 40000. I need to take a closer look to see what can be done to speed it up.
We could also do a "variant", where we run it twice, once with decompose=true and once decompose=false (like the variants for the SDPA tests for example).
Using slighly different solver options should fix most of the errors when the test problems are solved with COSMO.jl.
Current errors:
LPs
1e-6
it will need more iterations than the defaultmax_iter = 2500
SOCP
1e-5
, so slightly above the required solver tolerance, but accurate enough foratol
. I changed the MOI-wrapper to return the solution in such cases. (see commit: https://github.com/oxfordcontrol/COSMO.jl/commit/439f218cdfb0f6992d37d364e90917097d80ca53)SDPs
A = [B 0; 0 0] >= 0
whereB >= 0
. I am not sure atm how to efficiently check for these cases. That's why I turned the decomposition off for this testset, i.e.decomp = false
.5e-7
andmax_iter = 40000
. I need to take a closer look to see what can be done to speed it up.